Link Wallet for Agents: Letting AI Spend Money Without Handing Over Your Card

What Link Wallet for Agents Actually Is

Link Wallet for Agents is a new product from Stripe's Link that gives an AI agent a way to spend money on your behalf without ever seeing your real card details. The pitch is simple: your payment credentials stay hidden, and you approve every purchase before it goes through.

The core problem it addresses is a real one. As people build agents that can browse, book, and buy things, the payment step has been the awkward gap. Handing an autonomous process your credit card number is a bad idea. Link's answer is to sit between the agent and the merchant, issuing one-time-use cards or shared payment tokens so the agent can transact at businesses across the internet without holding anything sensitive.

According to the page, getting started involves pointing your agent to a `skill.md` file, authenticating, and granting access. It lists compatibility with Claude, OpenClaw, and custom agents. There's also a "View on GitHub" link, which suggests some of the integration surface is open for developers to inspect and wire up themselves.

How the Approval Flow Works

The design leans heavily on human-in-the-loop control, which is the most important part for anyone building real automations. When your agent is ready to buy something, you get a real-time notification. The Link app shows you a spend request with the details: the merchant, the amount, and what's being purchased. The example on the page shows an agent named "openclaw-jane" requesting $180 for a queen room in New Orleans through a merchant called Triplo, with Decline and Approve buttons.

Every transaction gets logged in a purchase history view, so you can see the merchant, timestamp, amount, status, payment method, and which agent made the request. That audit trail matters. If you're going to let software spend money, being able to reconstruct exactly what happened and who initiated it is table stakes.

There's a set of features listed as "coming soon" that hint at where this is headed. Granular agent controls would let you set rules for when an agent can spend with or without your approval, which is the difference between approving every $5 purchase manually and letting small routine buys go through automatically. Also promised: more payment methods including "next-generation payment protocols and digital currencies," and saved buying preferences so an agent can access your purchase history to shop smarter. Those are unshipped, so treat them as direction rather than fact.

Why This Matters for Automation Builders

If you're building agentic workflows, the payment layer has been a hard wall. You can get an agent to research flights, compare hotels, or reorder supplies, but the moment it needs to actually pay, you either stop and do it yourself or expose credentials in ways you shouldn't. Link is trying to make that last step safe and delegated rather than a manual handoff.

The key detail is that approval is required per purchase today. That's a sensible default and it means the wallet is not yet a fully autonomous spending machine. It's closer to giving your agent a debit request button that pings you for a signature. For high-value or one-off purchases, that friction is a feature. For high-frequency, low-value automations, you'll want those granular controls before this replaces manual buying, and those aren't here yet.

It's also worth being clear about what this is not. Link Wallet for Agents is a payment authorization and issuance layer, not an orchestration platform. It doesn't build your agent or decide what to buy. You still need the agent itself, and tools like Zapier, Make, or n8n if you're stitching together the surrounding workflow, plus a model provider like OpenAI or Claude to drive the reasoning. Link slots in specifically at the point of purchase.

Compared to the alternative today, which is mostly virtual card services or manual checkout, the differentiator is the agent-native design: the `skill.md` integration path, the per-agent identity in the approval and history views, and the explicit intent to support agent-specific spending rules. Whether it becomes the standard depends on merchant coverage and how those controls actually work once shipped.

A few honest caveats. The compatibility list is short, and OpenClaw is not a widely known name, so real-world coverage of "businesses across the internet" is a claim worth testing before you rely on it. The most useful features for automation, particularly unattended low-value spend, are still on the roadmap. And because this is a Stripe product, expect it to work best inside that ecosystem.

Still, the fundamentals are the right ones: credentials never exposed, per-transaction approval, and a clear audit trail. That's a reasonable foundation for anyone who wants an agent to spend money without losing sleep over it.

If you want help wiring an agent payment flow into your own automations, browse the provider directory to find people who can put it to work.